Predicting Dialogue Outcomes over Structured Latent Representations

Studying dialogue relations as a complex structured output prediction problem, independently of the dialogue outcome, has been the focus of significant research efforts, such as discourse relations [9], rhetorical structure [8, 1] and dialogue act modeling [13]. While these works capture linguisticallymotivated fine-grained dialogue relations, they require considerable effort and linguistic knowledge. Furthermore, the subjective nature of our task, focusing on dialogue relations relevant to the dialogue outcome, allow for many different interpretations of the interactions leading to the observed outcome.